Image Quality Assessment Based On Transfer Learning

With the rapid development of image technology,many scholars have paid much attention to the research of image quality evaluation and have made good progress.However,most of the algorithms focus on the full reference and reduced reference,which are more or less dependent on the reference image information.In this paper,a novel framework of no-reference image quality evaluation algorithm based on hierarchical attenuation characteristics is proposed,analyze the quality of images and the attenuation of images to predict image quality and feature migration method is used to extract the characteristics of image noise intensity and noise type,and the final non reference image quality evaluation is realized by using joint feature method.The research content of this article is divided into three parts:The first part constructs the level attenuation characteristic image quality evaluation.The analysis of human visual perception shows that human understanding of images is a process of hierarchical construction.First,we extract the local structure information of the image,establish the recognition of image contour,and then fuse the local structure to create image understanding and transform it into cognition and memory.Considering the human's primary visual neuron unit,we choose the mechanism of the direction selection mechanism to establish the local structural features,and extract the high-level semantics of the image using deep learning hierarchical network structure.The two are fused,and the image level features are established.The noise of different intensities also attenuates the image.When the noise is small,the image is affected by the local structure,and the noise greatly destroys the image's semantics.The attenuation characteristics of the hierarchical features are used to evaluate the quality of the image very well.The second part constructs the image quality evaluation based on the feature migration.Image noise has a connection to the quality of the image,and the image noise level is very difficult to divide.People's cognition of things is also a process of continuous learning.Deep learning has achieved good results in many visual tasks.Through deep learning network,the relationship between image noise level and image quality can be established,and we can learn the characteristics of image noise sensitive.In this paper,feature migration is used to extract the characteristics that are sensitive to the image noise level.The third part constructs the image quality evaluation based on the joint feature.Although the image quality assessment algorithm based on deep learning image has achieved good results,but in practical applications,it is easy to overmatch because of lack of enough training samples.In this paper,the effect of noise on image quality is analyzed,which shows that different noises have different influence on the image,and it is not enough to consider the image noise intensity unilaterally.This paper proposes an image quality evaluation algorithm based on joint feature on the basis of image level attenuation characteristics.Combined with the characteristics of directional selection,noise intensity and noise type,no reference image quality evaluation system is completed.The performance of the algorithm in the existing database shows that the image quality assessment algorithm based on migration learning presented in this paper has a strong consistency with human subjective perception.